Method and apparatus for graph data processing

Abstract

The application provides a method and device for graph data processing, and relates to the technical field of artificial intelligence, and the method comprises: a method for graph data processing, which comprises the following steps: acquiring graph data, extracting the basic structure features of the graph data through a static topology path, adaptively screening the key neighbor features in the graph data based on an attention mechanism and in combination with the basic structure features through a dynamic path, establishing long-distance semantic association between nodes through topology-aware coding based on the basic structure features and the key neighbor features through a linear graph transformer with O ( n ) of linear complexity, and establishing long-distance semantic association between nodes through topology-aware coding; performing frequency domain projection on the node features to generate frequency domain features, explicitly extracting and enhancing the periodic discriminant components in the frequency domain features through a Fourier analysis network to generate graph neural network data. Thus, the graph representation learning capability of the graph neural network is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and apparatus for graph data processing. Background Technology

[0002] Graph data naturally exists in complex systems such as social networks, biomolecules, and transportation systems. Its non-Euclidean properties (irregular relationships between nodes and dynamic topological evolution) make it difficult for traditional deep learning models (such as CNNs and RNNs) to directly model it. Graph neural networks, through message passing mechanisms, achieve node representation learning and iteratively aggregate local neighborhood information, becoming the core paradigm for processing graph data. Graph neural networks are frequently applied in key areas and practical scenarios such as social network analysis, molecular modeling and disease prediction, fraud detection and risk assessment, path optimization and traffic prediction. In the field of social network analysis, graph nodes represent individuals in a social network, and graph edges represent the social connections between nodes. In the field of molecular modeling, graph nodes represent the microscopic building blocks of molecules, and graph edges represent the chemical bonds between nodes. In the field of disease prediction, graph nodes can be categorized into different types based on the research scenario, such as independent research objects like genes, proteins, metabolites, cells, clinical samples (patients), and disease types, with graph edges representing biological interactions or connections between nodes. In the field of fraud detection, graph nodes represent various identified objects in financial or business scenarios, with graph edges representing the association between abnormal or normal business behaviors between nodes. In the field of risk assessment, graph nodes represent various subjects and objects associated with risk, with graph edges representing the risk association between nodes. In the field of path optimization, graph nodes represent geographical or functional nodes in transportation, logistics, or network systems, with graph edges representing the connectivity between nodes. In the field of traffic prediction, graph nodes represent traffic nodes in transportation, communication, or e-commerce scenarios, with graph edges representing the traffic transmission relationships between nodes.

[0003] Traditional graph neural networks (GNNs) suffer from low graph representation learning capabilities. When faced with graphs containing long-tailed nodes or excessively sparse nodes, traditional GNNs struggle to capture high-order structural information. Furthermore, Transformer-based graph models are inefficient and prone to neglecting local features. This theoretical limitation directly hinders the models' ability to achieve a qualitative breakthrough in graph representation learning when dealing with complex network data, failing to accurately reflect the full complexity of the graph structure.

[0004] Therefore, there is an urgent need for a method that can effectively improve the graph representation learning ability of graph neural networks. Summary of the Invention

[0005] To address the technical problem of low graph representation learning ability in existing graph neural networks, embodiments of the present invention provide a method and apparatus for graph data processing. The technical solution is as follows: On the one hand, a method for graph data processing is provided, the method comprising: Acquire graph data, extract the basic structural features of the graph data through static topological paths, and adaptively filter key neighbor features in the graph data based on an attention mechanism using dynamic paths, combined with the basic structural features. Based on the aforementioned basic structural features and key neighbor features, a linear graph transformer is used to... O ( n With linear complexity, long-distance semantic associations between nodes are established through topology-aware encoding; Node features are projected into the frequency domain to generate frequency domain features. The periodic discriminant components in the frequency domain features are explicitly extracted and enhanced through a Fourier analysis network to generate graph neural network data.

[0006] Optionally, the basic structural features and the key neighbor features are fused through residual connections.

[0007] Optionally, the linear graph transformer performs attention calculation using the following formula: ,in This is a random feature mapping function that maps high-dimensional features to a low-dimensional space. The dimension after mapping, and , where Q is the query matrix, K is the key matrix, and V is the value matrix.

[0008] Optionally, the linear graph transformer approximates the Gaussian kernel function using random Fourier features.

[0009] Optionally, frequency domain features are generated by frequency domain projection of the node features, including: Perform a Fast Fourier Transform on the node features to generate frequency domain features.

[0010] Optionally, the step of explicitly extracting and enhancing the periodic discriminant components in the frequency domain features through a Fourier analysis network to generate graph neural network data includes: By using a task-adaptive frequency domain filtering mechanism in a Fourier analysis network, useful periodic components are enhanced, noise is suppressed, and a frequency domain spectrum is generated. The frequency domain spectrum is mapped back to the spatial domain through inverse fast Fourier transform to generate graph neural network data.

[0011] Optionally, the Fourier analysis network is implemented using the following formula: ,in, Let be the learnable weight coefficients for the j-th frequency band. Weights of learnable complex-valued basis functions ( ), This represents element-wise multiplication. Indicates the total number of frequency bands. This indicates the inverse fast Fourier transform operation. This indicates the Fast Fourier Transform operation.

[0012] On the other hand, an apparatus for graph data processing is also provided, which is used to implement the graph data processing method provided in the embodiments of the present invention, the apparatus comprising: The graph context attention module is used to acquire graph data, extract the basic structural features of the graph data through static topological paths, and adaptively filter key neighbor features in the graph data based on attention mechanisms and the basic structural features through dynamic paths. The linear graph transformation module is used to transform the basic structural features and the key neighbor features using a linear graph transformer. O ( n With linear complexity, long-distance semantic associations between nodes are established through topology-aware encoding; The Fourier analysis network module is used to project node features into the frequency domain to generate frequency domain features. The Fourier analysis network is used to explicitly extract and enhance the periodic discriminant components in the frequency domain features to generate graph neural network data.

[0013] On the other hand, an apparatus for graph data processing is also provided, the apparatus for graph data processing comprising: processor; A memory storing computer-readable instructions, which, when executed by the processor, implement the method provided in the embodiments of the present invention.

[0014] On the other hand, embodiments of the present invention also provide a computer-readable storage medium storing program code, which can be invoked by a processor to execute the method provided in the embodiments of the present invention.

[0015] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: This invention, through the acquisition of graph data, extracts the basic structural features of the graph data using static topological paths, and adaptively filters key neighbor features in the graph data using dynamic paths based on an attention mechanism. Based on the basic structural features and the key neighbor features, a linear graph transformer is used to... O ( nThe method achieves linear complexity by establishing long-distance semantic associations between nodes through topologically aware encoding. It projects node features into the frequency domain to generate frequency domain features, and explicitly extracts and enhances the periodic discriminant components in these features using Fourier analysis networks to generate graph neural network data. Evolving graph neural network data through this method explicitly enhances the periodic features in node attributes, suppresses high-frequency noise, improves classification robustness, and enhances the generalization ability to periodic patterns. Thus, by improving the frequency domain perspective, it effectively enhances the representation learning ability of graph neural networks. In practical applications, the method provided in this invention can accurately capture periodic patterns in traffic flow and user behavior to improve prediction accuracy; effectively filter high-frequency noise in industrial sensor or financial transaction data to enhance anti-interference capabilities; and highlight the global community structure in social networks, thereby achieving better robustness and accuracy in tasks such as prediction, risk control, and anomaly detection. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of a method for graph data processing provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a device structure for graph data processing provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of a device structure for graph data processing provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0019] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0020] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0021] In this embodiment of the invention, sometimes a subscript such as W1 may be mistakenly written as a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0022] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0023] To address the technical problems of weak explicit modeling capabilities and high computational time complexity in existing graph neural network representation learning methods, this invention provides a method and apparatus for graph data processing. The technical solution is as follows: On the one hand, such as Figure 1 As shown, a method for graph data processing is provided, the method comprising: S1. Obtain graph data, extract the basic structural features of the graph data through static topological paths, and adaptively filter key neighbor features in the graph data based on attention mechanisms and the basic structural features through dynamic paths. S2. Based on the aforementioned basic structural features and key neighbor features, a linear graph transformer is used to... O ( n With linear complexity, long-distance semantic associations between nodes are established through topology-aware encoding; S3. Project the node features into the frequency domain to generate frequency domain features. Explicitly extract and enhance the periodic discriminant components in the frequency domain features through a Fourier analysis network to generate graph neural network data.

[0024] In some embodiments, this invention constructs a three-level serial collaborative architecture, achieving efficient representation learning of complex graph structures through end-to-end modeling of local feature enhancement, global semantic propagation, and frequency domain feature optimization. The three-level serial collaborative architecture includes a graph context attention module, a linear graph transformer, and a Fourier analysis network module.

[0025] The graph context attention module in the architecture extracts basic structural features through static topological paths, preserving core structural information for sparse nodes. Meanwhile, the dynamic path adaptively selects key neighbor features based on the attention mechanism. The residual fusion of the dual paths effectively solves the problem of local structural degradation and ensures semantic integrity during feature transmission.

[0026] The linear graph transformer inherits the output features of the graph context attention module, breaking through the O(n²) efficiency bottleneck of traditional transformer models with a linear complexity of O(n). It establishes long-distance semantic associations between nodes through topology-aware encoding, solving the efficiency dilemma of global dependency modeling on large-scale graphs and realizing efficient propagation from local features to global semantics.

[0027] The Fourier analysis network module, as the final feature enhancement layer, amplifies the periodic discriminative components in node features through explicit frequency domain decomposition, while suppressing high-frequency noise interference. This significantly improves the class discriminative power of long-tail nodes, solves the problem of periodic features being submerged by non-periodic noise, and provides a more robust feature representation for node classification tasks.

[0028] In some implementations, the core design goal of the static topology path in the graph context attention module is to provide stable basic structural features, especially for sparse nodes in long-tailed distributions. This is achieved by combining graph convolution operations with residual connections to suppress noise interference and preserve the basic topology information of the nodes. ,in, This represents the adjacency matrix after adding self-loops ( , This is the original adjacency matrix. (the identity matrix) Normalized degree matrix ( ), This is the GeLU activation function. The input is the node feature matrix.

[0029] The core of this formula lies in the symmetric normalization operation ( This balances the feature contributions of nodes with different degrees, avoiding excessive dominance of high-degree nodes in feature aggregation; simultaneously, the introduction of self-loops ( This ensures that each node retains its own characteristics during the aggregation process, effectively alleviating the problem of feature loss caused by insufficient sampling of sparse neighborhood nodes.

[0030] Residual joins are implemented as follows: , This is the GeLU activation function. The original features are those that are directly transmitted. After being added to the graph convolution result and then passed through an activation function, the gradient vanishing problem in deep networks is avoided, thus enhancing the stability of model training.

[0031] Dynamic Path aims to address the problem of insufficient utilization of local structural information by adaptively weighting features through an attention mechanism. Its core is to learn node-level attention weights by fusing static and original features, thereby achieving dynamic focusing on key features. Attention weight calculation: Attention weights The calculation follows the formula: ,in Representing static path characteristics With original features The concatenated vectors are used; a multilayer perceptron (including dimensionality reduction) is used to learn attention information from the concatenated features; a softmax function is used to normalize the attention scores. This design learns importance weights through feature self-interaction, avoiding information bias from a single feature source and ensuring that key features receive sufficient attention. Dynamic path output: After weighting the value features with attention weights, the output of the dynamic path is: ,in Value features obtained through graph convolution ( ), This represents element-wise multiplication, using attention weights. right Dynamic adjustments are made to enhance key features. Finally, residual fusion with static path features is performed. This approach balances structural stability and feature adaptability, enhancing the model's ability to model complex graph structures.

[0032] To balance the structural stability of static paths with the feature adaptability of dynamic paths, the graph context attention module employs residual connections to achieve efficient fusion of features from both paths. The specific fusion method is shown in the formula: ,in This is the static path output (node ​​neighborhood structure features extracted by GCN). The output is a dynamic path (adaptive features weighted by attention). Its core logic lies in this: static paths provide the "basic anchor points" of the graph topology, ensuring that the feature aggregation process does not deviate from the local structural relationships of nodes; dynamic paths, on the other hand, use an attention mechanism to "dynamically calibrate" key features, compensating for the shortcomings of static convolution in capturing complex semantic relationships. By directly superimposing dual-path features through residual connections, adaptive adjustment capabilities can be injected while preserving structural information, avoiding information bias from single-path features.

[0033] Furthermore, the fusion operation does not require the introduction of additional learnable parameters; feature fusion is achieved simply by adding elements one by one. This ensures the model's expressive power while controlling computational complexity, enabling the module to efficiently adapt to large-scale graph data.

[0034] In global dependency modeling of large-scale graphs, the global attention mechanism of traditional transformer-type models (such as Graphormer) faces a serious efficiency bottleneck: its computational complexity of O(n²) increases quadratically with the number of nodes n, making training on graphs with billions of nodes difficult. To overcome this predicament, this invention proposes to use a linear graph transformer, employing a dual strategy of kernel function approximation and matrix multiplication associative law optimization to reduce the global propagation complexity to the linear level of O(n).

[0035] The linear graph transformer restructures the computation process into a "mapping then aggregation" approach using the associative law of matrix multiplication. The restructured attention calculation formula is as follows: ,in For random feature mapping functions, high-dimensional features are mapped to a low-dimensional space. The dimension after mapping, and (), where Q is the query matrix, K is the key matrix, and V is the value matrix. After reconstruction, the computational complexity is reduced from... Down to :when When taking 512, a graph with billions of nodes ( The computational cost can be reduced by four orders of magnitude (from Down to ).

[0036] The core principle of this optimization is to utilize the associative law of matrix multiplication ( Change the computation order: through low-complexity feature mapping ( , )and Local aggregation ( This operation avoids the computational bottleneck of large-scale matrix multiplication.

[0037] The kernel function approximation is a further reduction of the mapping function. To reduce computational cost, the linear graph transformer uses a stochastic Fourier feature approximation of the Gaussian kernel function, with the mapping form as follows: ,in For a single node's feature vector, The weight vector is a randomly sampled vector (following a Gaussian distribution). , (This is the kernel width parameter, used to control the smoothness of the kernel function). This is a normalization factor to ensure the stability of the variance of the mapped features.

[0038] The theoretical basis for this approximation is Bochner's theorem: any positive definite kernel function can be expressed as the expectation of its characteristic functions. The Gaussian kernel, as a typical positive definite kernel, has characteristic functions that can be expressed as a linear combination of sine and cosine functions. Through random Fourier features, this can be achieved in low-dimensional space (…). Similarity calculations of approximately high-dimensional Gaussian kernels in (dimensional) are performed, thus avoiding explicit construction. Attention matrix – memory usage from traditional methods Reduce to That is, only the traditional method .

[0039] The linear graph transformer reduces the complexity to linear level through kernel function approximation and associative law optimization. .

[0040] In terms of reasoning speed, in the case of Benchmark tests on the OGBN-products dataset show that the linear graph transformer is 141 times faster than Graphormer (the linear graph transformer takes about 0.8 seconds per round of inference, while Graphormer takes about 112.8 seconds per round), significantly reducing the deployment cost and inference latency of large-scale graph models.

[0041] Fourier Analysis Network Module: In graph node classification tasks, node attributes often contain rich periodic features (such as the diurnal flow fluctuations of traffic networks and the weekly user activity patterns of social networks). However, traditional neural networks (such as multilayer perceptrons and transformers) have three inherent defects when modeling these features, which seriously affect model performance.

[0042] Black-box memory limitation: When faced with periodic signals, traditional models tend to mechanically memorize periodic patterns from training data through a large number of parameters, rather than learning the essence of periodic regularity. This results in poor extrapolation ability for unseen periodic intervals. The core reason for this limitation is that traditional models lack explicit modeling mechanisms for periodic features and cannot capture the inherent regularity of periodic signals.

[0043] Feature confusion problem: Periodic components in node attributes (such as the quarterly fluctuation features of economic networks) are often submerged by non-periodic noise (such as sudden transaction interference). Traditional linear layers use implicit feature mapping, which cannot explicitly decouple the discriminative components in the frequency domain, resulting in useful periodic features being masked by noise and affecting classification accuracy.

[0044] The parameter complexity of the fully connected layer in a traditional multilayer perceptron is... ( (for feature dimension), when With over a million parameters, deployment on large-scale graphs (>10 nodes) faces severe memory pressure, limiting the model's practical application scenarios.

[0045] This invention replaces the traditional multilayer perceptron layer with a Fourier analysis network module and solves these problems through frequency domain feature enhancement.

[0046] The Fourier analysis network module explicitly enhances periodic discriminative features through frequency domain basis function learning and adaptive filtering mechanisms, providing a lightweight and interpretable feature enhancement scheme for graph node classification and effectively solving these problems.

[0047] Extending the Fourier transform to graph structures, this applies to node feature vectors. ( (where is the feature dimension), its graphical Fourier transform is defined as: ,in For the graph Laplace matrix eigenvector matrix (satisfying) , (e.g., eigenvalue diagonal matrix) It is a frequency domain spectrum vector. is the spatial node feature vector. Its physical meaning is to project the node features in the spatial domain onto the orthogonal basis space in the frequency domain, so that the frequency characteristics (such as smoothness and periodicity) of the graph signal can be explicitly represented in the frequency domain. The low-frequency components correspond to the global smoothness features in the graph, and the high-frequency components correspond to local fluctuations or noise.

[0048] Fourier analysis network modules use learnable parameters By dynamically adjusting the weights of the frequency domain basis functions, adaptive frequency band selection can be achieved. The core transformation formula is as follows: ,in For spatial node feature vectors, The node feature vectors are enhanced by the Fourier analysis network. , , For learnable complex-valued basis function weights, For the first Learnable weighting coefficients for each frequency band Indicates the total number of frequency bands. This represents element-wise multiplication, and the Fourier analysis network module does not require pre-computation of the Laplacian eigenvectors (avoiding...). (Feature decomposition cost), significantly improving the practicality and efficiency of the model.

[0049] The Fourier analysis network module achieves explicit decoupling and enhancement of periodic features through a three-step process of "frequency domain projection - filtering enhancement - spatial domain reconstruction". The algorithm design is as follows: Frequency domain projection and component decoupling: for input spatial domain node feature vectors Perform a Fast Fourier Transform (FFT) to map spatial features to the frequency domain: ,in This indicates the Fast Fourier Transform operation. Given the feature vectors of spatial nodes, the output frequency domain spectrum is... (The real part is the cosine component, and the imaginary part is the sine component). The core function of this step is to decompose the periodic feature in the time domain into components of different frequencies, thereby achieving a preliminary separation between the periodic feature and noise (low frequencies correspond to periodic components, and high frequencies correspond to noise).

[0050] Learnable bandpass filtering: Designing task-adaptive frequency domain filtering mechanisms to enhance useful periodic components and suppress noise. ,in This represents the frequency domain spectrum vector obtained after the node features undergo a forward Fourier transform. Weights of learnable complex-valued basis functions ( ), This represents element-wise multiplication. Indicates the total number of frequency bands. Indicates the first The frequency domain spectrum vector after band filtering. The real part controls the retention intensity of the cosine component, and the imaginary part controls the retention intensity of the sine component. This is adaptively adjusted through training to match the frequency domain characteristics of the task; simultaneously, a scalar gain coefficient is introduced. Weighted enhancement of key periodic components: ,in Represents the final frequency domain spectrum vector. Indicates the first Frequency domain spectrum vector after band filtering For the first Learnable weighting coefficients for each frequency band The total number of frequency bands is represented by the core principle of this design: to dynamically adjust the frequency domain filtering strategy through learnable parameters, enabling the model to autonomously discover and enhance the periodic features most useful for classification tasks.

[0051] Inverse transform reconstruction enhances features: The filtered frequency domain spectrum is mapped back to the spatial domain using an inverse fast Fourier transform to obtain the enhanced features. ,in This indicates the inverse fast Fourier transform operation. This represents the node feature vector enhanced by the Fourier analysis network. Combining the above three steps, the complete mathematical expression of the Fourier analysis network module is: ,in For the first Learnable weighting coefficients for each frequency band Weights of learnable complex-valued basis functions ( ), Indicates the total number of frequency bands. The formula, which represents element-wise multiplication, clearly demonstrates the core mechanism by which the Fourier analysis network module enhances periodic features through frequency domain filtering.

[0052] The Fourier analysis network module in this embodiment of the invention, through explicit frequency domain decomposition and basis function learning, can understand the intrinsic laws of periodic characteristics. Its core formula is: Traditional GFT relies on a fixed Laplacian function ( (Assuming a predefined eigenvector matrix), its transformation formula is: It cannot adapt to the frequency domain requirements of different tasks. The Fourier analysis network module, through learnable... Dynamically adjust the frequency band weights.

[0053] Fourier analysis network module through Weight the critical periodic components and suppress high-frequency noise.

[0054] Regarding the number of parameters in lightweight structural design, the number of parameters in the Fourier analysis network module is: ( , (where J is the feature dimension), and J represents the total number of frequency bands. Representing the node feature dimension, in With this configuration, the number of parameters is only 4096, which is less than that of a multi-layer perceptron. The number of parameters is reduced by 62%, significantly alleviating the deployment pressure of large-scale graphs. In terms of computational efficiency, the frequency domain operation complexity of this embodiment is... More layered perceptrons Complexity in The computing speed is increased by 1.8 times, and with The increase in size makes the advantages more obvious. (This improves performance by 3.2 times), providing technical support for real-time inference of large-scale graphs.

[0055] Optionally, the basic structural features and the key neighbor features are fused through residual connections.

[0056] Optionally, the linear graph transformer performs attention calculation using the following formula: ,in This is a random feature mapping function that maps high-dimensional features to a low-dimensional space. The dimension after mapping, and , where Q is the query matrix, K is the key matrix, and V is the value matrix.

[0057] Optionally, the linear graph transformer approximates the Gaussian kernel function using random Fourier features.

[0058] Optionally, frequency domain features are generated by frequency domain projection of the node features, including: Perform a Fast Fourier Transform on the node features to generate frequency domain features.

[0059] Optionally, the step of explicitly extracting and enhancing the periodic discriminant components in the frequency domain features through a Fourier analysis network to generate graph neural network data includes: By using a task-adaptive frequency domain filtering mechanism in a Fourier analysis network, useful periodic components are enhanced, noise is suppressed, and a frequency domain spectrum is generated. The frequency domain spectrum is mapped back to the spatial domain through inverse fast Fourier transform to generate graph neural network data.

[0060] Optionally, the Fourier analysis network is implemented using the following formula: ,in, Let be the learnable weight coefficients for the j-th frequency band. Weights of learnable complex-valued basis functions ( ), This represents element-wise multiplication. Indicates the total number of frequency bands. This indicates the inverse fast Fourier transform operation. This indicates the Fast Fourier Transform operation.

[0061] On the other hand, such as Figure 2 As shown, an apparatus for graph data processing is also provided, which is used to implement the graph data processing method provided in the embodiments of the present invention. The apparatus includes: The graph context attention module 201 is used to acquire graph data, extract the basic structural features of the graph data through static topological paths, and adaptively filter key neighbor features in the graph data based on an attention mechanism through dynamic paths. Linear graph transformation module 202 is used to transform the basic structural features and the key neighbor features using a linear graph transformer. O ( n With linear complexity, long-distance semantic associations between nodes are established through topology-aware encoding; Fourier analysis network module 203 is used to project node features into the frequency domain to generate frequency domain features. The Fourier analysis network is used to explicitly extract and enhance the periodic discriminant components in the frequency domain features to generate graph neural network data.

[0062] On the other hand, an apparatus for graph data processing is also provided, the apparatus for graph data processing comprising: processor; A memory storing computer-readable instructions, which, when executed by the processor, implement the method provided in the embodiments of the present invention.

[0063] On the other hand, embodiments of the present invention also provide a computer-readable storage medium storing program code, which can be invoked by a processor to execute the method provided in the embodiments of the present invention.

[0064] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: This invention, through the acquisition of graph data, extracts the basic structural features of the graph data using static topological paths, and adaptively filters key neighbor features in the graph data using dynamic paths based on an attention mechanism. Based on the basic structural features and the key neighbor features, a linear graph transformer is used to... O ( n The method achieves linear complexity by establishing long-distance semantic associations between nodes through topologically aware encoding. It projects node features into the frequency domain to generate frequency domain features, and explicitly extracts and enhances the periodic discriminant components in these features using Fourier analysis networks to generate graph neural network data. Evolving graph neural network data through this method explicitly enhances the periodic features in node attributes, suppresses high-frequency noise, improves classification robustness, and enhances the generalization ability to periodic patterns. Thus, by improving the frequency domain perspective, it effectively enhances the representation learning ability of graph neural networks. In practical applications, the method provided in this invention can accurately capture periodic patterns in traffic flow and user behavior to improve prediction accuracy; effectively filter high-frequency noise in industrial sensor or financial transaction data to enhance anti-interference capabilities; and highlight the global community structure in social networks, thereby achieving better robustness and accuracy in tasks such as prediction, risk control, and anomaly detection.

[0065] Figure 3 This is a schematic diagram of the structure of a device for graph data processing provided in an embodiment of the present invention, such as... Figure 3 As shown, optionally, the device 310 for graph data processing may include a first processor 2001.

[0066] Optionally, the device 310 for graph data processing may also include a memory 2002 and a transceiver 2003.

[0067] The first processor 2001, memory 2002, and transceiver 2003 can be connected via a communication bus.

[0068] The following is combined with Figure 3 A detailed description of each component of the device 310 used for graph data processing is provided below: The first processor 2001 is the control center of the device 310 for graph data processing. It can be a single processor or a collective term for multiple processing elements. For example, the first processor 2001 can be one or more central processing units (CPUs), application-specific integrated circuits (ASICs), or one or more integrated circuits configured to implement embodiments of the present invention, such as one or more digital signal processors (DSPs), or one or more field-programmable gate arrays (FPGAs).

[0069] Optionally, the first processor 2001 can perform various functions of the device 310 for graph data processing by running or executing software programs stored in the memory 2002 and calling data stored in the memory 2002.

[0070] In a specific implementation, as one example, the first processor 2001 may include one or more CPUs, for example... Figure 3 CPU0 and CPU1 are shown in the diagram.

[0071] In a specific implementation, as one example, the device 310 for graph data processing may also include multiple processors, for example... Figure 3 The first processor 2001 and the second processor 2004 are shown in the diagram. Each of these processors can be a single-core processor or a multi-core processor. Here, a processor can refer to one or more devices, circuits, and / or processing cores used to process data (such as computer program instructions).

[0072] The memory 2002 is used to store the software program that executes the present invention, and is controlled by the first processor 2001 to execute it. The specific implementation method can be referred to the above method embodiment, and will not be repeated here.

[0073] Optionally, the memory 2002 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code having an instruction or data structure form and accessible by a computer, but not limited thereto. The memory 2002 may be integrated with the first processor 2001 or may exist independently, and may be connected via an interface circuit of the device 310 for graph data processing. Figure 3 (Not shown in the image) is coupled to the first processor 2001, and this embodiment of the invention does not specifically limit this.

[0074] The transceiver 2003 is used to communicate with network devices or with terminal devices.

[0075] Alternatively, transceiver 2003 may include a receiver and a transmitter. Figure 3 (Not shown separately). The receiver is used to implement the receiving function, and the transmitter is used to implement the transmitting function.

[0076] Alternatively, the transceiver 2003 can be integrated with the first processor 2001 or exist independently, and can be connected via the interface circuit of the device 310 for graph data processing. Figure 3 (Not shown in the image) is coupled to the first processor 2001, and this embodiment of the invention does not specifically limit this.

[0077] It should be noted that, Figure 3 The structure of the device 310 for graph data processing shown in the diagram does not constitute a limitation on the router. Actual knowledge structure identification devices may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0078] Furthermore, the technical effects of the device 310 for graph data processing can be referred to the technical effects of the method for graph data processing described in the above method embodiments, and will not be repeated here.

[0079] It should be understood that the first processor 2001 in the embodiments of the present invention may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0080] It should also be understood that the memory in the embodiments of the present invention can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0081] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, motor drive, or data center to another website, computer, motor drive, or data center via infrared, microwave, or other means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a motor drive or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0082] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0083] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0084] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0085] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0086] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0087] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0088] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0089] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0090] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a motor driver, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0091] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for graph data processing, characterized in that, The method includes: Acquire graph data, extract the basic structural features of the graph data through static topological paths, and adaptively filter key neighbor features in the graph data based on an attention mechanism using dynamic paths, combined with the basic structural features. Based on the aforementioned basic structural features and key neighbor features, a linear graph transformer is used to... O ( n With linear complexity, long-distance semantic associations between nodes are established through topology-aware encoding; Node features are projected into the frequency domain to generate frequency domain features. The periodic discriminant components in the frequency domain features are explicitly extracted and enhanced through a Fourier analysis network to generate graph neural network data.

2. The method according to claim 1, characterized in that, The basic structural features and the key neighbor features are fused through residual connections.

3. The method according to claim 1, characterized in that, The linear graph transformer calculates attention using the following formula: ,in This is a random feature mapping function that maps high-dimensional features to a low-dimensional space. The dimension after mapping, and , where Q is the query matrix, K is the key matrix, and V is the value matrix.

4. The method according to claim 3, characterized in that, The linear graph transformer approximates the Gaussian kernel function using random Fourier features.

5. The method according to claim 1, characterized in that, Frequency domain features are generated by projecting node features into the frequency domain, including: Perform a Fast Fourier Transform on the node features to generate frequency domain features.

6. The method according to claim 1, characterized in that, The process of explicitly extracting and enhancing the periodic discriminant components in the frequency domain features using a Fourier analysis network to generate graph neural network data includes: By using a task-adaptive frequency domain filtering mechanism in a Fourier analysis network, useful periodic components are enhanced, noise is suppressed, and a frequency domain spectrum is generated. The frequency domain spectrum is mapped back to the spatial domain through inverse fast Fourier transform to generate graph neural network data.

7. The method according to claim 1, characterized in that, The Fourier analysis network is implemented using the following formula: ,in, Let be the learnable weight coefficients for the j-th frequency band. Weights of learnable complex-valued basis functions ( ), This represents element-wise multiplication. Indicates the total number of frequency bands. This indicates the inverse fast Fourier transform operation. This indicates the Fast Fourier Transform operation.

8. An apparatus for graph data processing, the apparatus for graph data processing being used to implement the method for graph data processing as described in any one of claims 1-7, characterized in that, The device includes: The graph context attention module is used to acquire graph data, extract the basic structural features of the graph data through static topological paths, and adaptively filter key neighbor features in the graph data based on attention mechanisms and the basic structural features through dynamic paths. The linear graph transformation module is used to transform the basic structural features and the key neighbor features using a linear graph transformer. O ( n With linear complexity, long-distance semantic associations between nodes are established through topology-aware encoding; The Fourier analysis network module is used to project node features into the frequency domain to generate frequency domain features. The Fourier analysis network is used to explicitly extract and enhance the periodic discriminant components in the frequency domain features to generate graph neural network data.

9. An apparatus for graph data processing, characterized in that, The device for graph data processing includes: processor; A memory storing computer-readable instructions that, when executed by the processor, implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium contains program code that can be invoked by a processor to execute the method as described in any one of claims 1 to 7.

Images